1 Lecture 18: Large Caches, Multiprocessors Today: NUCA caches, multiprocessors (Sections ) Reminder: assignment 5 due Thursday (don’t procrastinate!)

Core 0 L1 D$ L1 I$ L2 $ Core 1 L1 D$ L1 I$ L2 $ Core 2 L1 D$ L1 I$ L2 $ Core 3 L1 D$ L1 I$ L2 $ Core 4 L1 D$ L1 I$ L2 $ Core 5 L1 D$ L1 I$ L2 $ Core 6 L1 D$ L1 I$ L2 $ Core 7 L1 D$ L1 I$ L2 $ Memory Controller for off-chip access A single tile composed of a core, L1 caches, and a bank (slice) of the shared L2 cache The cache controller forwards address requests to the appropriate L2 bank and handles coherence operations Distributed Shared Cache

3 The L2 (or L3) can be a large shared cache, but is physically partitioned into banks and distributed on chip Each core (tile) has one L2 cache bank adjacent to it One bank stores a subset of “sets” and all ways for that set OS-based first-touch page coloring can force a thread’s pages to have physical page numbers that map to the thread’s local L2 bank Physical Address | color | Physical page # Cache Index

4 UCA and NUCA The small-sized caches so far have all been uniform cache access: the latency for any access is a constant, no matter where data is found For a large multi-megabyte cache, it is expensive to limit access time by the worst case delay: hence, non-uniform cache architecture The distributed shared cache is an example of a NUCA cache: variable latency to each bank

5 NUCA Design Space Distribute Sets: Static-NUCA: Each block has a unique location; easy to find data; page coloring for locality; page migration if initial mapping is sub-optimal Distribute Ways: Dynamic-NUCA: More flexibility in block placement; complicated search mechanisms; blocks migrate to be closer to their accessor Private data are easy to handle; Shared data must be placed at the center-of-gravity of accesses

6 Prefetching Hardware prefetching can be employed for any of the cache levels It can introduce cache pollution – prefetched data is often placed in a separate prefetch buffer to avoid pollution – this buffer must be looked up in parallel with the cache access Aggressive prefetching increases “coverage”, but leads to a reduction in “accuracy”  wasted memory bandwidth Prefetches must be timely: they must be issued sufficiently in advance to hide the latency, but not too early (to avoid pollution and eviction before use)

7 Stream Buffers Simplest form of prefetch: on every miss, bring in multiple cache lines When you read the top of the queue, bring in the next line L1 Stream buffer Sequential lines

8 Stride-Based Prefetching For each load, keep track of the last address accessed by the load and a possibly consistent stride FSM detects consistent stride and issues prefetches init trans steady no-pred incorrect correct incorrect (update stride) correct incorrect (update stride) incorrect (update stride) tagprev_addrstridestate PC

9 Taxonomy SISD: single instruction and single data stream: uniprocessor MISD: no commercial multiprocessor: imagine data going through a pipeline of execution engines SIMD: vector architectures: lower flexibility MIMD: most multiprocessors today: easy to construct with off-the-shelf computers, most flexibility

10 Memory Organization - I Centralized shared-memory multiprocessor or Symmetric shared-memory multiprocessor (SMP) Multiple processors connected to a single centralized memory – since all processors see the same memory organization  uniform memory access (UMA) Shared-memory because all processors can access the entire memory address space Can centralized memory emerge as a bandwidth bottleneck? – not if you have large caches and employ fewer than a dozen processors

11 SMPs or Centralized Shared-Memory Processor Caches Processor Caches Processor Caches Processor Caches Main Memory I/O System

12 Memory Organization - II For higher scalability, memory is distributed among processors  distributed memory multiprocessors If one processor can directly address the memory local to another processor, the address space is shared  distributed shared-memory (DSM) multiprocessor If memories are strictly local, we need messages to communicate data  cluster of computers or multicomputers Non-uniform memory architecture (NUMA) since local memory has lower latency than remote memory

13 Distributed Memory Multiprocessors Processor & Caches MemoryI/O Processor & Caches MemoryI/O Processor & Caches MemoryI/O Processor & Caches MemoryI/O Interconnection network

14 Shared-Memory Vs. Message-Passing Shared-memory: Well-understood programming model Communication is implicit and hardware handles protection Hardware-controlled caching Message-passing: No cache coherence  simpler hardware Explicit communication  easier for the programmer to restructure code Sender can initiate data transfer

15 Ocean Kernel Procedure Solve(A) begin diff = done = 0; while (!done) do diff = 0; for i  1 to n do for j  1 to n do temp = A[i,j]; A[i,j]  0.2 * (A[i,j] + neighbors); diff += abs(A[i,j] – temp); end for if (diff < TOL) then done = 1; end while end procedure

16 Shared Address Space Model int n, nprocs; float **A, diff; LOCKDEC(diff_lock); BARDEC(bar1); main() begin read(n); read(nprocs); A  G_MALLOC(); initialize (A); CREATE (nprocs,Solve,A); WAIT_FOR_END (nprocs); end main procedure Solve(A) int i, j, pid, done=0; float temp, mydiff=0; int mymin = 1 + (pid * n/procs); int mymax = mymin + n/nprocs -1; while (!done) do mydiff = diff = 0; BARRIER(bar1,nprocs); for i  mymin to mymax for j  1 to n do … endfor LOCK(diff_lock); diff += mydiff; UNLOCK(diff_lock); BARRIER (bar1, nprocs); if (diff < TOL) then done = 1; BARRIER (bar1, nprocs); endwhile

17 Message Passing Model main() read(n); read(nprocs); CREATE (nprocs-1, Solve); Solve(); WAIT_FOR_END (nprocs-1); procedure Solve() int i, j, pid, nn = n/nprocs, done=0; float temp, tempdiff, mydiff = 0; myA  malloc(…) initialize(myA); while (!done) do mydiff = 0; if (pid != 0) SEND(&myA[1,0], n, pid-1, ROW); if (pid != nprocs-1) SEND(&myA[nn,0], n, pid+1, ROW); if (pid != 0) RECEIVE(&myA[0,0], n, pid-1, ROW); if (pid != nprocs-1) RECEIVE(&myA[nn+1,0], n, pid+1, ROW); for i  1 to nn do for j  1 to n do … endfor if (pid != 0) SEND(mydiff, 1, 0, DIFF); RECEIVE(done, 1, 0, DONE); else for i  1 to nprocs-1 do RECEIVE(tempdiff, 1, *, DIFF); mydiff += tempdiff; endfor if (mydiff < TOL) done = 1; for i  1 to nprocs-1 do SEND(done, 1, I, DONE); endfor endif endwhile

18 Title Bullet